AI-powered meeting intelligence for a $400 Billion wealth platform
Advisors used to record meetings with third-party apps. The notes stayed outside the platform: disconnected from the CRM, invisible to compliance, re-entered by hand. SteelSprint shipped a working demo in one month and production in four.
The Company
A publicly traded Canadian wealth management firm. $400 billion in assets under management. Thousands of advisors across North America.
The Problem
Advisors recorded client meetings with third-party apps. The notes stayed outside the platform: disconnected from the CRM, invisible to compliance, re-entered by hand after every meeting.
The firm wanted to bring that capability in-house. Record a meeting, extract structured data, push it into the wealth platform. They had the product vision but not the ML engineering team to build it. Hiring would take longer than building.
The Approach
We split the work into two phases. The split was the strategy.
A standalone frontend: live recording, real-time transcription, and structured extraction. No backend, no CRM integration. A product you could use in a room and see results.
When development is fast enough, you don't pitch a concept. You build it. Stakeholders react to working software, not slides. That's a different conversation.
CTO approved after the demo. We integrated the meeting intelligence module into the firm's existing wealth platform.
We designed the AI layer to be vendor-flexible from the start. Swapping providers meant changing configuration, not rebuilding the product. In financial services, that kind of flexibility isn't optional.
Technical detailsPhase 1 shipped a standalone demo using real-time transcription and structured extraction. Phase 2 integrated into the wealth platform with a vendor-flexible AI layer that moved between providers without rewriting extraction logic.+
React, TypeScript, Tailwind. Real-time transcription via Whisper. Structured extraction via GPT-4.1. LLMs wrote most of the Phase 1 code. Humans steered architecture and product decisions. ESLint and TypeScript enforced consistency at write time.
Fastify backend, Kysely for data access, tRPC for type-safe client-server communication. The LLM layer moved from OpenAI to AWS Bedrock running Claude Sonnet 4.5.
The AI extraction pipeline was built against an abstracted model interface from the start. Switching from OpenAI to Bedrock required configuration changes only, with no rewrites to the extraction or summarization logic.
What We Delivered
A meeting intelligence module integrated into the firm's wealth platform:
Live audio capture during advisor-client meetings
Real-time transcription of meeting audio into text
Structured extraction: summaries, action items, and client details pulled automatically from the transcript
Platform integration: extracted data flows directly into existing client records
Vendor-flexible AI: the AI provider can be swapped without rebuilding the product
The Results
One month to a working demo. Three more to production. Four months from kickoff to a shipped feature in a national wealth platform.
The demo replaced months of evaluation. Working software moved the CTO's decision in weeks. Phase 2 delivered the full module: extraction, platform integration, and production deployment. The feature entered phased rollout to advisors.
The Takeaway
When development is fast enough to replace a pitch with a product, the buy-in cycle compresses. The thing you evaluate is the thing you ship.